BioChirp Help & Documentation

General LLMs answer directly from their internal training data and, in some cases, from simple web or RAG search. BioChirp, in contrast, treats curated databases (TTD, CTD, HCDT) as the primary source of truth. The LLM is used mainly to understand your query, normalize terms, map them onto database schemas, set dynamic similarity cut-offs, and summarize the resulting tables. This “database-first, LLM-last” design reduces hallucinations and makes answers easier to audit.

BioChirp minimizes hallucinations by grounding all core answers in structured database rows. Final summaries are generated from the “head view” (top 50 rows) of the complete results table, so context-length limitations rarely apply.

However, minor variations can still occur, mainly in the synonym and family member expansion step. BioChirp uses three complementary processes:

• Fuzzy search: Always returns reproducible results for a given input.
• Official APIs (e.g., PubChem, HGNC): Provide consistent results across runs, as long as the external API services are available and unchanged.
• Semantic transformer embeddings: This method uses dynamic similarity thresholds and an LLM-based false-positive filter, which can occasionally introduce minor differences in the set of returned synonyms or family members between runs.

These factors can sometimes lead to small, non-deterministic changes in the complete association list, but overall, BioChirp remains much more consistent and auditable than standard LLM-based tools.

BioChirp is directly integrated with several authoritative, curated biomedical databases, including:

• TTD (Therapeutic Target Database): Comprehensive resource for drug–target–disease relationships and target classes.
• CTD (Comparative Toxicogenomics Database): Chemical–gene–disease interactions and information on environmental exposures.
• HCDT (Highly Confident Drug–Target Database): High-confidence drug–target associations with strong supporting evidence.

When relevant associations are not found in these databases, BioChirp can supplement answers using trusted web and biomedical literature sources, which are always clearly labeled and kept separate from database results.

BioChirp restricts where the LLM can introduce uncertainty by enforcing a strict structure. The Interpreter and Planner must generate dictionaries and join-plans that match the underlying database schema; if they fail to map cleanly, the system does not guess and instead requests clarification or routes the query elsewhere.

For entity expansion, BioChirp combines fuzzy search and deep-learning embeddings similarity, along with LLM based selection, with results from official biomedical APIs (e.g., PubChem, HGNC/HUGO, OpenTargets) to retrieve standardized synonyms and family members. These expansions are then filtered with dynamic similarity thresholds so only well-supported candidates are included.

Even with synonym/family expansion, the final associations always come directly from curated databases. If no rows are found or if the query falls outside BioChirp’s database scope, the system will not fabricate relations—it clearly returns “no result.”

When database results are unavailable, BioChirp falls back to web and literature search using trusted biomedical sources through its Tavily-like search layer. This ensures the answer remains grounded in reputable, verifiable resources rather than random web pages.

BioChirp currently operates on offline, locally downloaded snapshots of curated biomedical databases. We plan to update these datasets periodically—typically once a year or as needed based on major upstream releases or user requirements.

Any significant updates to the underlying data or system will be announced so users remain aware of version changes and improvements.

BioChirp allows you to export results in CSV or JSON formats when the query falls within the scope of the available biomedical data. The columns and entity names are designed to be stable, making it easy to integrate the data into your own analysis pipelines.

If you use BioChirp's services, please be sure to cite us as part of your work to acknowledge the platform and its data.

BioChirp is designed to operate without requiring user logins, ensuring that no user-identifiable data is stored. User privacy is prioritized, and we do not collect any personal information during interactions.